[Beowulf] [owner-chemistry at ccl.net: CCL:G: parallel G03 in Linux cluster]

[Reuti]

Quoting Mark Hahn <hahn at physics.mcmaster.ca>:

>> >> The OpenMP parallelization should also work without Linda - but of course
>> >> limited to some routines and 2 cores.
>
> I wonder if this is only true if you, for instance, compile gaussian
> with a PGI demo license or something.  also, the normal (old) way of
> using SMP is fork-based parallelism with SYSV shm.  is this separate
> (orthogonal) to OpenMP-based parallelism?  I suspect the latter is only
> used for SMP parallelism within parallelized blas libraries like MKL.

I compiled G03.C02 without OpenMP and tried to use it still with 
%nprocs but it
was not working, as it was before in G03.B05. So I contacted Gaussian and they
stated, that the old fork-model is gone since C.02 (or maybe even since C.01).
The SMP is now only OpenMP according to this info I have.

>
>> machines with single dual-core CPUs. Let's rephrase: ...but limited to the
>> cores in each of the machines.
>
> mosix, I believe, provides network-shared SYSV shm.  I wonder if anyone has

You mean like e.g. http://www.kerrighed.org , indeed, would be interesting.

-- Reuti


> tried to use gaussian that way.  depending on the sharing patters, it might
> work pretty well - my understanding is that the shm is used mainly for
> caching intermediate values (integrals?), which might be
> write-once-read-mostly, and therefore friendly to the kind of caching
> necessary for any page-based net-shm implementation.
>
> regards, mark hahn.
>